Controlled-Source Audio Magnetotelluric (CSAMT) is a near-surface geophysical method that developed on the basis of Magnetotelluric method (MT). With the development of social economy,the data quality of CSAMT has also been seriously disturbed by noise interference. In practical exploration,the time series of electromagnetic field is usually superimposed with large-scale trend drift,short-term sudden strong interference and peak impulsive outliers,resulting in the distortion of the calculated resistivity spectrum. In this paper,an anti-interference processing method based on deep learning and joint de-noising is proposed to preprocess CSAMT time series. Firstly,a forward algorithm of electromagnetic time series of layered earth controllable source is proposed,which is used to generate standard electromagnetic signals without noise interference. Then,a Long and Short Term Memory Neural Network (LSTM) classifier is trained to recognize the noise. Finally,the improved Empirical Mode Decomposition (EMD) algorithm,correlation based data selection algorithm and robust statistical algorithm are jointly used to de-noise the CSAMT time series. The test results by simulated data show that the recognition accuracy of LSTM for noise interference can reach more than 95%,and the three noise reduction algorithms can reduce the data error from about 20% to less than 3%. Finally,the proposed method is applied to the actual data set of a metal mining area in Inner Mongolia. the accuracy of low-frequency resistivity and phase was effectively improved.